Moving from drought hazard to impact forecasts

Present-day drought early warning systems provide the end-users information on the ongoing and forecasted drought hazard (e.g. river flow deficit). However, information on the forecasted drought impacts, which is a prerequisite for drought management, is still missing. Here we present the first study assessing the feasibility of forecasting drought impacts, using machine-learning to relate forecasted hydro-meteorological drought indices to reported drought impacts. Results show that models, which were built with more than 50 months of reported drought impacts, are able to forecast drought impacts a few months ahead. This study highlights the importance of drought impact databases for developing drought impact functions. Our findings recommend that institutions that provide operational drought early warnings should not only forecast drought hazard, but also impacts after developing an impact database.


Supplementary Figures
Supplementary Figure 1. Germany divided into NUTS-1 regions, including their acronyms.
When forecasting droughts, it would be beneficial for end-users and stakeholders to have more detailed information about possible impacts. Details on temporal and spatial scale and specific type of impact are, however, strongly related to the impact data available. To find a balance between reasonable forecasts and the level of detail, 5 cases were identified. The temporal reso-

Workflow
In As already discussed in Section 3.3, the EDII inventory was translated into binary time series of drought impacts. An overview of these binary time series is shown in

Supplementary Notes
Analyses in this study were carried out based on the German NUTS-1 regions. The NUTS regions are geo-coded standard regions developed by the European Union 2 . In Germany, the NUTS-1 regions correspond with the federal states ( Supplementary Fig. 1). Figure 1 shows the German NUTS-1 regions and their acronyms.
Collection of drought impact reports and their entries in the EDII 1 is still an ongoing process. The text-based reports are obtained from various sources, such as governmental reports, NGO reports, newspapers, digital media, and scientific papers. Impact entries in the EDII must contain: (1) temporal reference that can be indicated by month, season, or year; and (2) spatial reference, which is the location of the reported impacts. This can be either referred to different levels of geographical regions using the European Union NUTS standard (Nomenclature of Units for Territorial Statistics) or specified by rivers and lakes. The drought impact reports are divided into 15 impact categories: (1) agriculture and livestock farming, (2) forestry, (3) freshwater aquaculture and fisheries, (4) energy and industry, (5) waterborne transportation, (6) tourism and recreation, (7) public water supply, (8) water quality, (9) freshwater ecosystems, (10) terrestrial ecosystem, (11) soil system, (12) wildfires, (13) air quality, (14) human health and public safety, and (15) conflicts ( Supplementary Fig. 2). Supplementary Figure 2 shows that in Germany most drought impacts occurred during the summer period and mainly in the southern and western regions (BV, BW, and BB, see Supplementary Fig. 1 for the acronyms). The most frequent reported drought impacts are in the categories agriculture and livestock farming, freshwater ecosystems, public water supply, and water-borne transportation.
Availability of reported impact information in the EDII appears to vary across Europe. This results in temporal and spatial biases within the EDII. For the present study, Germany was selected to explore the potential of drought impact forecasting, as it is one of the most documented countries within the EDII.
An example of time series of drought events in the German NUTS-1 regions RP and BB is presented in Supplementary Figure 3. These events were calculated using proxy observed data for Standardized Precipitation Index and Standardized Runoff Index accumulated over 6 months (SPI-6 and SRI-6, respectively). The optimal accumulation period for standardized drought indices depends on catchment characteristics (fast versus slowly-responding catchments), but also on the impacted sector. For some sectors, which largely depend on soil moisture, an accumulation period of 3 months (SPI-3) fits well, for other sectors that are more influenced by groundwater, or groundwater-fed rivers, longer accumulation periods are selected (e.g. SPI-6). For instance, the heat maps compiled by Ref. 3 show that accumulation periods over 6 months (SPI-x, x>6) are typical for groundwater. The temporal evolution of the SPI-6 and SRI-6 shows that droughts occurred in RP and BB in 2003, 2006, and 2008.
Drought impact forecasting functions have been developed for the German NUTS-1 regions with sufficient impact data for the drought impact groups: (1) agriculture and livestock farming, and forestry; (2) energy and industry, waterborne transportation, and public water supply; (3) water quality, freshwater ecosystem, and terrestrial ecosystem; and (4) wildfire, air quality, and human health and public safety. These functions were trained using observed drought events derived from drought indices (SPI-x, SPEI-x, and SRI-x, with x=1, 3, 6, 12), years in drought, and months in drought from 1990 to 2017 as predictors, and binary drought impact time series. Supplementary Figure 4 illustrates the predictor importance for every region and for selected merged impact categories (impact groups), which are generated with the Random Forest algorithm. The dark gray boxes indicate that insufficient data are available for a certain impact group and NUTS-1 region to develop drought impact forecasting functions.
Overall, the year of impact occurrence seems to be a good predictor for all the impact groups (most right bar in each histogram). This is the result of temporal bias within the EDII. For impact Group 1, accumulation periods of 3 and 6 months are best linked with the impact (Supplementary Fig. 4a). SPEI and SRI are better drought impact predictors than the SPI. The accumulation period of 6 months appears to be the best predictor for impact Group 2, closely followed by the accumulation periods of 3 and 12 months ( Supplementary Fig. 4b). Drought indices SPEI and SRI are best linked to this impact group. As this category is well covered across Germany, spatial trends can be found. For the more southern regions, accumulation periods of 3 and 6 months are best linked. The northern regions are well correlated with longer accumulation periods (6 and 12 months). Supplementary Figure 4c shows that best predictors for impact Group 3 have accumulation periods of 3 and 6 months. Like the other two impact groups, the drought indices SPEI and SRI have better predictive power than SPI. No clear spatial trends can be found across regions for this impact group. For impacts related to fires, health, and air quality (Group 4), predictors with short accumulation period are most important for forecasting this drought impact, closely followed by predictors with accumulation periods of 6 months ( Supplementary Fig. 4d).
In our study, we used methods that are already well established. As described in the Method section, data and methods used in this study are similar with studies described in Ref. 4, 5, 6, and 7. Supplementary Figure 5 shows the flowchart describing the data and methods.